Optimizing Artificial Neural Network-Based Models to Predict Rice Blast Epidemics in Korea.

Area of Science:

• Agricultural science
• Computational biology
• Plant pathology

Background:

• Accurate rice blast prediction is crucial for crop management.
• Machine learning (ML) methods offer potential for disease forecasting.
• Data quality and quantity significantly impact ML model performance.

Purpose of the Study:

• To develop and compare artificial neural network (ANN)-based models for predicting rice blast.
• To evaluate the impact of diverse input datasets on model performance.
• To identify the optimal ANN architecture and data combination for rice blast prediction.

Main Methods:

• Developed three ANN models: Feed-Forward Neural Network (FFNN) and Long Short-Term Memory (LSTM).
• Combined FFNN and LSTM with different input datasets, including historical blast occurrence and weather data (temperature, humidity, precipitation).
• Compared model performance using recall scores, focusing on the Blast_Weather_FFNN model.

Main Results:

• The Blast_Weather_FFNN model achieved the highest recall score of 66.3% for rice blast prediction.
• This model utilized 3 years of historical blast occurrence data and January-July weather data.
• Model performance improved with optimized hyperparameter tuning and appropriate machine learning techniques.

Conclusions:

• Artificial neural network models, particularly FFNN combined with specific historical and weather data, can effectively predict rice blast.
• Optimizing input data and hyperparameter tuning are critical for enhancing ANN-based disease prediction models.
• Systematic collection of long-term disease data is essential for improving agricultural disease forecasting accuracy.